1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25 /*
26 * Copyright (c) 2012 by Delphix. All rights reserved.
27 */
28
29 #include <sys/zfs_context.h>
30 #include <sys/spa.h>
31 #include <sys/dmu.h>
32 #include <sys/zio.h>
33 #include <sys/space_map.h>
34
35 static kmem_cache_t *space_seg_cache;
36
37 void
38 space_map_init(void)
39 {
40 ASSERT(space_seg_cache == NULL);
41 space_seg_cache = kmem_cache_create("space_seg_cache",
42 sizeof (space_seg_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
43 }
44
45 void
46 space_map_fini(void)
47 {
48 kmem_cache_destroy(space_seg_cache);
49 space_seg_cache = NULL;
50 }
51
52 /*
53 * Space map routines.
54 * NOTE: caller is responsible for all locking.
55 */
56 static int
57 space_map_seg_compare(const void *x1, const void *x2)
58 {
59 const space_seg_t *s1 = x1;
60 const space_seg_t *s2 = x2;
61
62 if (s1->ss_start < s2->ss_start) {
63 if (s1->ss_end > s2->ss_start)
64 return (0);
65 return (-1);
66 }
67 if (s1->ss_start > s2->ss_start) {
68 if (s1->ss_start < s2->ss_end)
69 return (0);
70 return (1);
71 }
72 return (0);
73 }
74
75 void
76 space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift,
77 kmutex_t *lp)
78 {
79 bzero(sm, sizeof (*sm));
80
81 cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL);
82
83 avl_create(&sm->sm_root, space_map_seg_compare,
84 sizeof (space_seg_t), offsetof(struct space_seg, ss_node));
85
86 sm->sm_start = start;
87 sm->sm_size = size;
88 sm->sm_shift = shift;
89 sm->sm_lock = lp;
90 }
91
92 void
93 space_map_destroy(space_map_t *sm)
94 {
95 ASSERT(!sm->sm_loaded && !sm->sm_loading);
96 VERIFY0(sm->sm_space);
97 avl_destroy(&sm->sm_root);
98 cv_destroy(&sm->sm_load_cv);
99 }
100
101 void
102 space_map_add(space_map_t *sm, uint64_t start, uint64_t size)
103 {
104 avl_index_t where;
105 space_seg_t *ss_before, *ss_after, *ss;
106 uint64_t end = start + size;
107 int merge_before, merge_after;
108
109 ASSERT(MUTEX_HELD(sm->sm_lock));
110 VERIFY(!sm->sm_condensing);
111 VERIFY(size != 0);
112 VERIFY3U(start, >=, sm->sm_start);
113 VERIFY3U(end, <=, sm->sm_start + sm->sm_size);
114 VERIFY(sm->sm_space + size <= sm->sm_size);
115 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
116 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
117
118 ss = space_map_find(sm, start, size, &where);
119 if (ss != NULL) {
120 zfs_panic_recover("zfs: allocating allocated segment"
121 "(offset=%llu size=%llu)\n",
122 (longlong_t)start, (longlong_t)size);
123 return;
124 }
125
126 /* Make sure we don't overlap with either of our neighbors */
127 VERIFY(ss == NULL);
128
129 ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
130 ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
131
132 merge_before = (ss_before != NULL && ss_before->ss_end == start);
133 merge_after = (ss_after != NULL && ss_after->ss_start == end);
134
135 if (merge_before && merge_after) {
136 avl_remove(&sm->sm_root, ss_before);
137 if (sm->sm_pp_root) {
138 avl_remove(sm->sm_pp_root, ss_before);
139 avl_remove(sm->sm_pp_root, ss_after);
140 }
141 ss_after->ss_start = ss_before->ss_start;
142 kmem_cache_free(space_seg_cache, ss_before);
143 ss = ss_after;
144 } else if (merge_before) {
145 ss_before->ss_end = end;
146 if (sm->sm_pp_root)
147 avl_remove(sm->sm_pp_root, ss_before);
148 ss = ss_before;
149 } else if (merge_after) {
150 ss_after->ss_start = start;
151 if (sm->sm_pp_root)
152 avl_remove(sm->sm_pp_root, ss_after);
153 ss = ss_after;
154 } else {
155 ss = kmem_cache_alloc(space_seg_cache, KM_SLEEP);
156 ss->ss_start = start;
157 ss->ss_end = end;
158 avl_insert(&sm->sm_root, ss, where);
159 }
160
161 if (sm->sm_pp_root)
162 avl_add(sm->sm_pp_root, ss);
163
164 sm->sm_space += size;
165 }
166
167 void
168 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
169 {
170 avl_index_t where;
171 space_seg_t *ss, *newseg;
172 uint64_t end = start + size;
173 int left_over, right_over;
174
175 VERIFY(!sm->sm_condensing);
176 ss = space_map_find(sm, start, size, &where);
177
178 /* Make sure we completely overlap with someone */
179 if (ss == NULL) {
180 zfs_panic_recover("zfs: freeing free segment "
181 "(offset=%llu size=%llu)",
182 (longlong_t)start, (longlong_t)size);
183 return;
184 }
185 VERIFY3U(ss->ss_start, <=, start);
186 VERIFY3U(ss->ss_end, >=, end);
187 VERIFY(sm->sm_space - size <= sm->sm_size);
188
189 left_over = (ss->ss_start != start);
190 right_over = (ss->ss_end != end);
191
192 if (sm->sm_pp_root)
193 avl_remove(sm->sm_pp_root, ss);
194
195 if (left_over && right_over) {
196 newseg = kmem_cache_alloc(space_seg_cache, KM_SLEEP);
197 newseg->ss_start = end;
198 newseg->ss_end = ss->ss_end;
199 ss->ss_end = start;
200 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
201 if (sm->sm_pp_root)
202 avl_add(sm->sm_pp_root, newseg);
203 } else if (left_over) {
204 ss->ss_end = start;
205 } else if (right_over) {
206 ss->ss_start = end;
207 } else {
208 avl_remove(&sm->sm_root, ss);
209 kmem_cache_free(space_seg_cache, ss);
210 ss = NULL;
211 }
212
213 if (sm->sm_pp_root && ss != NULL)
214 avl_add(sm->sm_pp_root, ss);
215
216 sm->sm_space -= size;
217 }
218
219 space_seg_t *
220 space_map_find(space_map_t *sm, uint64_t start, uint64_t size,
221 avl_index_t *wherep)
222 {
223 space_seg_t ssearch, *ss;
224
225 ASSERT(MUTEX_HELD(sm->sm_lock));
226 VERIFY(size != 0);
227 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
228 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
229
230 ssearch.ss_start = start;
231 ssearch.ss_end = start + size;
232 ss = avl_find(&sm->sm_root, &ssearch, wherep);
233
234 if (ss != NULL && ss->ss_start <= start && ss->ss_end >= start + size)
235 return (ss);
236 return (NULL);
237 }
238
239 boolean_t
240 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
241 {
242 avl_index_t where;
243
244 return (space_map_find(sm, start, size, &where) != 0);
245 }
246
247 void
248 space_map_swap(space_map_t **msrc, space_map_t **mdst)
249 {
250 space_map_t *sm;
251
252 ASSERT(MUTEX_HELD((*msrc)->sm_lock));
253 ASSERT0((*mdst)->sm_space);
254 ASSERT0(avl_numnodes(&(*mdst)->sm_root));
255
256 sm = *msrc;
257 *msrc = *mdst;
258 *mdst = sm;
259 }
260
261 void
262 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
263 {
264 space_seg_t *ss;
265 void *cookie = NULL;
266
267 ASSERT(MUTEX_HELD(sm->sm_lock));
268
269 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
270 if (func != NULL)
271 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
272 kmem_cache_free(space_seg_cache, ss);
273 }
274 sm->sm_space = 0;
275 }
276
277 void
278 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
279 {
280 space_seg_t *ss;
281
282 ASSERT(MUTEX_HELD(sm->sm_lock));
283
284 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
285 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
286 }
287
288 /*
289 * Wait for any in-progress space_map_load() to complete.
290 */
291 void
292 space_map_load_wait(space_map_t *sm)
293 {
294 ASSERT(MUTEX_HELD(sm->sm_lock));
295
296 while (sm->sm_loading) {
297 ASSERT(!sm->sm_loaded);
298 cv_wait(&sm->sm_load_cv, sm->sm_lock);
299 }
300 }
301
302 /*
303 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
304 * The caller must be OK with this.
305 */
306 int
307 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
308 space_map_obj_t *smo, objset_t *os)
309 {
310 uint64_t *entry, *entry_map, *entry_map_end;
311 uint64_t bufsize, size, offset, end, space;
312 uint64_t mapstart = sm->sm_start;
313 int error = 0;
314
315 ASSERT(MUTEX_HELD(sm->sm_lock));
316 ASSERT(!sm->sm_loaded);
317 ASSERT(!sm->sm_loading);
318
319 sm->sm_loading = B_TRUE;
320 end = smo->smo_objsize;
321 space = smo->smo_alloc;
322
323 ASSERT(sm->sm_ops == NULL);
324 VERIFY0(sm->sm_space);
325
326 if (maptype == SM_FREE) {
327 space_map_add(sm, sm->sm_start, sm->sm_size);
328 space = sm->sm_size - space;
329 }
330
331 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
332 entry_map = zio_buf_alloc(bufsize);
333
334 mutex_exit(sm->sm_lock);
335 if (end > bufsize)
336 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
337 mutex_enter(sm->sm_lock);
338
339 for (offset = 0; offset < end; offset += bufsize) {
340 size = MIN(end - offset, bufsize);
341 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
342 VERIFY(size != 0);
343
344 dprintf("object=%llu offset=%llx size=%llx\n",
345 smo->smo_object, offset, size);
346
347 mutex_exit(sm->sm_lock);
348 error = dmu_read(os, smo->smo_object, offset, size, entry_map,
349 DMU_READ_PREFETCH);
350 mutex_enter(sm->sm_lock);
351 if (error != 0)
352 break;
353
354 entry_map_end = entry_map + (size / sizeof (uint64_t));
355 for (entry = entry_map; entry < entry_map_end; entry++) {
356 uint64_t e = *entry;
357
358 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
359 continue;
360
361 (SM_TYPE_DECODE(e) == maptype ?
362 space_map_add : space_map_remove)(sm,
363 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
364 SM_RUN_DECODE(e) << sm->sm_shift);
365 }
366 }
367
368 if (error == 0) {
369 VERIFY3U(sm->sm_space, ==, space);
370
371 sm->sm_loaded = B_TRUE;
372 sm->sm_ops = ops;
373 if (ops != NULL)
374 ops->smop_load(sm);
375 } else {
376 space_map_vacate(sm, NULL, NULL);
377 }
378
379 zio_buf_free(entry_map, bufsize);
380
381 sm->sm_loading = B_FALSE;
382
383 cv_broadcast(&sm->sm_load_cv);
384
385 return (error);
386 }
387
388 void
389 space_map_unload(space_map_t *sm)
390 {
391 ASSERT(MUTEX_HELD(sm->sm_lock));
392
393 if (sm->sm_loaded && sm->sm_ops != NULL)
394 sm->sm_ops->smop_unload(sm);
395
396 sm->sm_loaded = B_FALSE;
397 sm->sm_ops = NULL;
398
399 space_map_vacate(sm, NULL, NULL);
400 }
401
402 uint64_t
403 space_map_maxsize(space_map_t *sm)
404 {
405 ASSERT(sm->sm_ops != NULL);
406 return (sm->sm_ops->smop_max(sm));
407 }
408
409 uint64_t
410 space_map_alloc(space_map_t *sm, uint64_t size)
411 {
412 uint64_t start;
413
414 start = sm->sm_ops->smop_alloc(sm, size);
415 if (start != -1ULL)
416 space_map_remove(sm, start, size);
417 return (start);
418 }
419
420 void
421 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
422 {
423 sm->sm_ops->smop_claim(sm, start, size);
424 space_map_remove(sm, start, size);
425 }
426
427 void
428 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
429 {
430 space_map_add(sm, start, size);
431 sm->sm_ops->smop_free(sm, start, size);
432 }
433
434 /*
435 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
436 */
437 void
438 space_map_sync(space_map_t *sm, uint8_t maptype,
439 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
440 {
441 spa_t *spa = dmu_objset_spa(os);
442 avl_tree_t *t = &sm->sm_root;
443 space_seg_t *ss;
444 uint64_t bufsize, start, size, run_len, total, sm_space, nodes;
445 uint64_t *entry, *entry_map, *entry_map_end;
446
447 ASSERT(MUTEX_HELD(sm->sm_lock));
448
449 if (sm->sm_space == 0)
450 return;
451
452 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
453 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
454 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
455 sm->sm_space);
456
457 if (maptype == SM_ALLOC)
458 smo->smo_alloc += sm->sm_space;
459 else
460 smo->smo_alloc -= sm->sm_space;
461
462 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
463 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
464 entry_map = zio_buf_alloc(bufsize);
465 entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
466 entry = entry_map;
467
468 *entry++ = SM_DEBUG_ENCODE(1) |
469 SM_DEBUG_ACTION_ENCODE(maptype) |
470 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
471 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
472
473 total = 0;
474 nodes = avl_numnodes(&sm->sm_root);
475 sm_space = sm->sm_space;
476 for (ss = avl_first(t); ss != NULL; ss = AVL_NEXT(t, ss)) {
477 size = ss->ss_end - ss->ss_start;
478 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
479
480 total += size;
481 size >>= sm->sm_shift;
482
483 while (size) {
484 run_len = MIN(size, SM_RUN_MAX);
485
486 if (entry == entry_map_end) {
487 mutex_exit(sm->sm_lock);
488 dmu_write(os, smo->smo_object, smo->smo_objsize,
489 bufsize, entry_map, tx);
490 mutex_enter(sm->sm_lock);
491 smo->smo_objsize += bufsize;
492 entry = entry_map;
493 }
494
495 *entry++ = SM_OFFSET_ENCODE(start) |
496 SM_TYPE_ENCODE(maptype) |
497 SM_RUN_ENCODE(run_len);
498
499 start += run_len;
500 size -= run_len;
501 }
502 }
503
504 if (entry != entry_map) {
505 size = (entry - entry_map) * sizeof (uint64_t);
506 mutex_exit(sm->sm_lock);
507 dmu_write(os, smo->smo_object, smo->smo_objsize,
508 size, entry_map, tx);
509 mutex_enter(sm->sm_lock);
510 smo->smo_objsize += size;
511 }
512
513 /*
514 * Ensure that the space_map's accounting wasn't changed
515 * while we were in the middle of writing it out.
516 */
517 VERIFY3U(nodes, ==, avl_numnodes(&sm->sm_root));
518 VERIFY3U(sm->sm_space, ==, sm_space);
519 VERIFY3U(sm->sm_space, ==, total);
520
521 zio_buf_free(entry_map, bufsize);
522 }
523
524 void
525 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
526 {
527 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
528
529 smo->smo_objsize = 0;
530 smo->smo_alloc = 0;
531 }
532
533 /*
534 * Space map reference trees.
535 *
536 * A space map is a collection of integers. Every integer is either
537 * in the map, or it's not. A space map reference tree generalizes
538 * the idea: it allows its members to have arbitrary reference counts,
539 * as opposed to the implicit reference count of 0 or 1 in a space map.
540 * This representation comes in handy when computing the union or
541 * intersection of multiple space maps. For example, the union of
542 * N space maps is the subset of the reference tree with refcnt >= 1.
543 * The intersection of N space maps is the subset with refcnt >= N.
544 *
545 * [It's very much like a Fourier transform. Unions and intersections
546 * are hard to perform in the 'space map domain', so we convert the maps
547 * into the 'reference count domain', where it's trivial, then invert.]
548 *
549 * vdev_dtl_reassess() uses computations of this form to determine
550 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
551 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
552 * has an outage wherever refcnt >= vdev_children.
553 */
554 static int
555 space_map_ref_compare(const void *x1, const void *x2)
556 {
557 const space_ref_t *sr1 = x1;
558 const space_ref_t *sr2 = x2;
559
560 if (sr1->sr_offset < sr2->sr_offset)
561 return (-1);
562 if (sr1->sr_offset > sr2->sr_offset)
563 return (1);
564
565 if (sr1 < sr2)
566 return (-1);
567 if (sr1 > sr2)
568 return (1);
569
570 return (0);
571 }
572
573 void
574 space_map_ref_create(avl_tree_t *t)
575 {
576 avl_create(t, space_map_ref_compare,
577 sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
578 }
579
580 void
581 space_map_ref_destroy(avl_tree_t *t)
582 {
583 space_ref_t *sr;
584 void *cookie = NULL;
585
586 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
587 kmem_free(sr, sizeof (*sr));
588
589 avl_destroy(t);
590 }
591
592 static void
593 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
594 {
595 space_ref_t *sr;
596
597 sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
598 sr->sr_offset = offset;
599 sr->sr_refcnt = refcnt;
600
601 avl_add(t, sr);
602 }
603
604 void
605 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
606 int64_t refcnt)
607 {
608 space_map_ref_add_node(t, start, refcnt);
609 space_map_ref_add_node(t, end, -refcnt);
610 }
611
612 /*
613 * Convert (or add) a space map into a reference tree.
614 */
615 void
616 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
617 {
618 space_seg_t *ss;
619
620 ASSERT(MUTEX_HELD(sm->sm_lock));
621
622 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
623 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
624 }
625
626 /*
627 * Convert a reference tree into a space map. The space map will contain
628 * all members of the reference tree for which refcnt >= minref.
629 */
630 void
631 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
632 {
633 uint64_t start = -1ULL;
634 int64_t refcnt = 0;
635 space_ref_t *sr;
636
637 ASSERT(MUTEX_HELD(sm->sm_lock));
638
639 space_map_vacate(sm, NULL, NULL);
640
641 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
642 refcnt += sr->sr_refcnt;
643 if (refcnt >= minref) {
644 if (start == -1ULL) {
645 start = sr->sr_offset;
646 }
647 } else {
648 if (start != -1ULL) {
649 uint64_t end = sr->sr_offset;
650 ASSERT(start <= end);
651 if (end > start)
652 space_map_add(sm, start, end - start);
653 start = -1ULL;
654 }
655 }
656 }
657 ASSERT(refcnt == 0);
658 ASSERT(start == -1ULL);
659 }